[Sequela02]

Part 7 of the move to Effect became fixing the places in the engine's loop that do not hear a stop (debated with GPT-6 Astra and GPT-6.1 Sol, and measured; the owner approved fixing them directly rather than moving loop.ts to Effect). This is the permission question.

Defect

permitted() (packages/core/src/loop.ts) awaited permissions.ask(call, reason), which takes no signal, and abort() and dispose() await the run. Stopping a run while the operator was asked, and not answering, left abort() waiting until someone answered. Measured: a session whose ask never answers did not end 15 s after the stop.

Reachable through the engine's API (a host that asks and is stopped), not through today's shell: there, Esc with the question open goes to the question and answers no, and a background session's waiting question is refused when it closes.

Change

• The loop races the operator's answer against the run's stop (untilStopped): a stop ends the wait, records the call as refused ("The run was stopped before the operator answered."), and reports it as stopped. The question itself is left to settle on its own; Permissions.ask keeps its signature, since no host has a stop to pass it today.
• operatorDecision turns an answer into the decision and reason, which keeps permitted within the lint's complexity limit.

Checks

• New test in packages/core/tests/permissions.test.ts: a question never answered, a stop, and the run ends aborted, the call not run, the refusal recorded. On dev the run never ends (the test times out); with the change it passes.
• bun run check locally (Windows): root 264 passed; engine 332 passed, 3 skipped; lint and types clean. The budget fails on shell memory (122 MB, limit 115 MB) on this machine; CI measured 109.6 MB on refactor: run the gate's verification with Effect #487. CI decides.